OpenAI tests AI in industrial design
OpenAI extends GPT-6 Astra to device design, automotive, and mechanical engineering. Trials show gains in iterations, but manufacturing remains the real test.
On November 3, 2026, OpenAI presented a series of physical engineering projects carried out with GPT-6 Astra, a multimodal model capable of handling mechanical, electronic, and software design tasks. The company tested a device intended to communicate with Codex, expanded its work with Chip Ganassi Racing, and opened a research collaboration with Mazda. The goal is clear: to move AI from drawing to manufacturable parts without breaking the production chain.
When AI draws, simulates, and codes instead of teams
The most telling test concerns an experimental device designed to talk with Codex. According to OpenAI, Astra defined functional requirements, designed the casing and internal supports, selected electrical components, drew the printed circuit board, and generated the necessary software integrations. The model also produced the user interface, without direct modification by team members on the design outputs.
In an earlier demonstration, Astra had already placed components and routed copper connections in KiCad, a printed circuit board design software. Here, two people designed the device in less than a week, on a side project, then a subcontractor produced a small batch of prototypes in less than a month. The story seems fast. Industrial reality is much less so.
Manufacturing remains the ultimate test
OpenAI indicates that production preparation required several review and iteration cycles. Astra did not always integrate manufacturing constraints, tolerances, or subcontractor assembly requirements. These limits serve precisely to identify what future models will need to master better: generating a complete design is not enough; it must also be made industrializable.
This is where the subject goes beyond a simple demonstrator. In engineering, a good CAD file is not worth a finished product. Between the two lie material constraints, cost trade-offs, deadlines, and workshop feedback. It is often at this point that lab dreams meet the screwdrivers.
BenchCAD 2.0 aims to measure progress, not hype
OpenAI also says it uses its own hardware and infrastructure teams to test Codex and ChatGPT on CAD modeling, revisions, simulation setup, and design reviews. Model usage, measured in tokens, has increased 23-fold since February, according to the company. This figure reflects model consumption, not directly time saved or final product quality.
To evaluate design capability, OpenAI cites BenchCAD 2.0, an internal test bench that asks models to recreate 3D parts and assemblies, then identify electronic board connections. The company reports an average score of 0.875 for a new internal model, versus 0.594 for GPT-6 Astra in its highest reasoning setting. The comparison suggests clear progress on design tasks but does not cover the entire engineering chain.
At Chip Ganassi Racing and Mazda, AI operates under human supervision
OpenAI conducts this work with a small group of industrial partners whose engineers review model outputs to verify accuracy and practical use. The idea is not to let AI improvise alone but to identify tasks it can truly help accomplish and those that still need manual rework.
At Chip Ganassi Racing, the expanded collaboration tests workflows that transform 3D scans of existing parts into editable CAD models and 2D manufacturing drawings, especially when original files are missing. The team's engineers then validate the accuracy and manufacturability of the results. During the first year of the partnership, work focused mainly on sensor data, race data, and pit stop performance; the new phase adds parts design and prototyping.
With Mazda, OpenAI combines its research and Forward Deployed Engineering teams with aerodynamic and crash safety engineers. Simulations focus on airflow and structural behavior while considering style and cost. In the aerodynamic area, GPT-6 Astra generates several front hood shapes, runs flow simulations, then refines the most promising variants to reduce drag.
Akihiro Nakata, lead engineer at Mazda, explains that simulating dozens of variants helps understand which details really matter for aerodynamics and where designers have leeway. OpenAI says it plans to share more results later and also invites engineering companies and software publishers involved in design, simulation, manufacturing, and production to propose high-value use cases. The real question at this stage is less whether AI knows how to draw than when it stops stumbling over the constraints of the physical world.
Key points
- November 3, 2026: OpenAI announcement on GPT-6 Astra.
- Two people designed the prototype in less than a week.
- A small batch of prototypes was produced in less than a month.
- Model usage multiplied by 23 since February at OpenAI.
- BenchCAD 2.0 score: 0.875 versus 0.594 for GPT-6 Astra.
- Chip Ganassi Racing and Mazda among cited partners.
By the numbers
- 23 times — increase in model usage within hardware teams since February, according to OpenAI.
- 0.875 — average score of a new internal model on BenchCAD 2.0.
- 0.594 — average score of GPT-6 Astra on the same benchmark, highest reasoning setting.
- 2 people — team that designed the experimental device, according to OpenAI.
- less than a month — production time for a small batch of prototypes by the subcontractor.